net: don't constantly post notifications to clients until they send

task/net/src/main.rs

@@ -257,9 +257,17 @@ fn main() -> ! {
     // who was waiting to hear back on their TX queue becoming non-full will
     // snap out of it.
     //
-    // This only works because we've set waiting_to_send to true for all sockets
-    // above.
-    server.wake_sockets();
+    // Note that we don't, actually, need to check whether the TX queue is empty
+    // or not here. We just reset, so any previously buffered packets that we
+    // had in the TX queue for that socket are definitely gone and the queue
+    // will be empty.
+    //
+    // Once we have woken everyone once, we can then only wake tasks when we are
+    // sure they actually are trying to send.
+    for (task_id, notification) in generated::SOCKET_OWNERS {
+        let task_id = sys_refresh_task_id(task_id);
+        sys_post(task_id, notification);
+    }
 
     // Go!
     loop {

task/net/src/server.rs

@@ -570,9 +570,15 @@ where
 
         Self {
             eth,
-            // The 'true' here is load-bearing: it ensures that sockets receive
-            // a notification on stack restart.
-            client_waiting_to_send: [true; SOCKET_COUNT],
+            // This only gets set if a client task has *definitely* tried to
+            // send on its socket. We do one big all-call wake everyone as soon
+            // as the net task starts, in case they were previously trying to
+            // send but their queue was full. If we were to keep waking any time
+            // a task's tx queue was not full, we will spuriously wake that task
+            // over and over again until it actually decides to send something
+            // --- which, given that communication on the management network is
+            // not generally initiated by the SP, could be a while!
+            client_waiting_to_send: [false; SOCKET_COUNT],
             vlan_state: enum_map::EnumMap::from_array(
                 vlan_state.into_array().unwrap_lite(),
             ),